Predicting who is most at risk of having a heart attack or a stroke is a difficult problem. Doctors decide who should get treatment based on a range of factors such as their blood pressure, how much they smoke and their weight. Heart attacks and strokes are caused by a fatty build-up in the blood vessels around the heart, which develops as people get older. Doctors cannot see how much of this fatty build up there is without doing risky or expensive tests. In older people, who are most at risk of heart attacks and strokes, it is common to have x-ray scans to check that their bones are not becoming too thin. During these scans, any fatty build-up in the largest blood vessel, the aorta, can be seen. With a cheap and quick way to measure this fatty build-up, doctors would be better able to decide who should be getting medication to prevent heart attacks and strokes. If a computer can be taught to automatically do these measurements, then lots of this information could be accessed without taking up any additional time from doctors. By showing a computer enough examples of how to perform a task, the computer can learn patterns and apply these to new examples. In this project the computer is shown a large number of the x-ray images with the fatty build-ups highlighted and taught to highlight these areas on images that it hasn’t seen before.
Motivation:
Cardiovascular disease (CVD) is a broad class of diseases involving the heart or blood vessels; including some of the most common causes of death globally. The most common forms involve ischaemic heart diseases, such as myocardial infarction, and cerebrovascular diseases, such as stroke. The main process driving these diseases is the development of atherosclerosis, fatty plaque formation within artery walls that narrows the artery and disrupts blood flow. It has been established that these diseases are highly preventable, but a lack of clinical signs ahead of a major cardiovascular event hinders identification of at-risk individuals for intervention.
Abdominal aortic calcification (AAC) develops as a result of atherosclerosis in the aorta. Calcification occurs within the plaques and can further inhibit the normal function of the aorta. The severity of AAC has been found to be a predictor of future cardiovascular events, even when controlling for other known risk factors. AAC is highly correlated with other predictors of CVD, such as coronary artery calcification and acts as a measure of atherosclerotic extent within the arterial system. These factors have led to increased scientific interest in using AAC as a screening tool and incorporating it in current CVD risk assessment scores.
The extent of AAC can be visualised using a range of imaging modalities. In particular, dual energy x-ray absorptiometry (DXA), commonly used to monitor bone mineral density (BMD) in individuals at risk of osteoporosis, uses a low radiation dose and is sensitive to calcification. During vertebral fracture assessment (VFA), DXA scanners use a lateral projection to assess the vertebrae. As the abdominal aorta lies just anterior to the lumbar vertebrae, it has been shown that calcification in this area can be imaged incidentally. AAC can be quantified using a 24-point scoring system, based on the length of the calcifications relative to the height of the lumbar vertebrae.
Despite the benefits of AAC assessment and recommendations for its use in CVD risk stratification, it is still not routine for VFA reports to include more than qualitative comments on AAC. There is definite potential for image segmentation techniques to locate and measure calcification in the abdominal aorta and report this automatically. With accurate automatic scoring system, large volumes of VFA images could be quickly and consistently assessed for AAC, to aid in research and to improve clinical treatment decisions
Approach:
The current approach to the problem uses deep learning to segment the areas of the images containing using deep learning. VFA images are annotated for pixels that contain calcification to create mask images. Pairs of images and masks can then be used to train models to segment images it is yet to encounter. The most promising models so far for this problem have been convolutional neural networks (CNNs). CNNs have been used in a range of machine learning tasks, including classification and detection. Fully convolutional networks (FCNs) have been effective in a variety of semantic segmentation tasks, as they are able to output a labelling prediction in the form of a mask of the same resolution as the input image. The exact network architectures used can be found in the publications below.
The main shortcoming of deep learning is the requirement for large amount of labelled data in order to train the networks. In order to select only the relevant parts of the VFA images (the lumbar region) and to increase the amount of training data available, the approach uses statistical shape models. The spatial relationship between the lumbar vertebrae and the anterior and posterior walls of the abdominal aorta is encoded from sagittal CT slices into a shape model. This model can then be fitted to the lumbar vertebrae of the VFA images and used to predict the region of interest containing the abdominal aorta. Small amounts of noise applied to the vertebral landmark points create non-linear deformations in the abdominal area of the image. This technique can be used, along with small translation, scaling and rotational transformations to create multiple training examples from each original image mask pair, greatly increasing the available data and the segmentation accuracy of the network.
Once the FCN has been trained it can produce a segmentation mask for each test image. These segmentations can then be used to automatically calculate 24-point scores. The shape model already warps each level of the lumbar vertebrae into a consistent proportion of the image, and a calculation of the line that best separates the anterior and posterior calcifications can be calculated. This allows the calculation of the proportion of calcification in each segment relative to the height of the segment. This allows an AAC24 score to be produced for images that have not been used to train the network, and to compare these scores to those of human annotators. The segmentation accuracy of the current method exceeds that of previous approaches and the correlation between automated and manual AAC24 scoring is encouraging.
Publications:
Chaplin, L. and Cootes, T., 2019, March. Automated scoring of aortic calcification in vertebral fracture assessment images. In Medical Imaging 2019: Computer-Aided Diagnosis (Vol. 10950, p. 1095038). International Society for Optics and Photonics – Available here